In [1]:
import latent_space
from PIL import Image
import numpy as np
import output
from histology_main import get_session,close_session
from constants import FULL_OUTPUT_PATH, BASE_PATH, BATCH_SIZE, Z_NOISE_DIM
import tensorflow as tf
import matplotlib.pylab as plt
from matplotlib.pyplot import figure, imshow, axis
from matplotlib.image import imread
from os import listdir, remove
import random
from glob import glob
%matplotlib inline
import dcgan_alt as dcgan
from utils import change_active_config
from evaluation import load_graph, read_tensor_from_image_file, load_labels
In [2]:
def label_and_show(src_dir, model_file, labels_file, input_height = 299,input_width = 299, image_list=None, show=0):
input_mean = 0
input_std = 255
input_layer = "Placeholder"
output_layer = "final_result"
files_to_show=[]
if image_list:
[files_to_show.append(src_dir+"/"+file) for file in image_list]
else:
files = [os.path.join(src_dir, file) for file in listdir(src_dir)]
files_to_show = random.sample(files,show)
graph = load_graph(model_file)
fig = plt.figure(figsize=(30,20))
for i,f in enumerate(files_to_show):
t = read_tensor_from_image_file(
f,
input_height=input_height,
input_width=input_width,
input_mean=input_mean,
input_std=input_std)
input_name = "import/" + input_layer
output_name = "import/" + output_layer
input_operation = graph.get_operation_by_name(input_name)
output_operation = graph.get_operation_by_name(output_name)
with tf.Session(graph=graph) as sess:
results = sess.run(output_operation.outputs[0], {
input_operation.outputs[0]: t
})
results = np.squeeze(results)
predic = results.argsort()[::-1]
labels = load_labels(labels_file)
subplt=fig.add_subplot(np.sqrt(show), np.sqrt(show) ,i+1)
plt.xticks([])
plt.yticks([])
plt.tight_layout()
color="green" if (labels[predic[0]].startswith("ben") or labels[predic[0]].startswith("neg")) else "red"
subplt.set_xlabel(labels[predic[0]]+" "+str(results[predic[0]]),color=color).set_fontsize(20)
name = f.split("/")[-1]
color="green" if name.startswith("neg") else "red" if name.startswith("pos") else "blue"
subplt.set_title(name,color=color).set_fontsize(22)
image = imread(f)
imshow(image)
CelebA : Aritmetica de vectores conceptuales con imagenes histólogicas¶
In [3]:
dataset = "camelyon_17_16_level0"
change_active_config(dataset)
In [4]:
session, input_real, input_z = get_session(1)
z_batch_tensor = tf.random.uniform(
(BATCH_SIZE, Z_NOISE_DIM), dtype=tf.float32, minval=-1, maxval=1)
In [21]:
prefix = "histo_"
#clean up
[os.remove(file) for file in glob(src_dir_test+"*"+prefix+"*.*")]
#Generamos unas 3000 imagenes (32*100)
output.generate_samples(session, z_batch_tensor, input_z, BATCH_SIZE * 100, save_tensor=True, name_prefix=prefix)
In [78]:
all_files = glob(src_dir_test+"*"+prefix+"*.png")
files_to_show = random.sample(all_files,24)
plt.figure(figsize=(20, 12))
plt.subplots_adjust(hspace=0.12, wspace=0.03)
for i,img_file in enumerate(files_to_show):
img = plt.imread(img_file)
plt.subplot(4, 6, i+1).set_title(img_file.split("/")[-1]).set_fontsize(10)
plt.imshow(img, aspect="auto")
plt.xticks([])
plt.yticks([])
i=i+1
Seleccionamos 3 imágenes que nos parezcan claramente negativas y en los que se vea la luz de estructuras tubulares, otras 3 que no contengan ninguna de estas estructuras y 3 más, cancerosas en este caso y también sin estructuras tubulares. Vamos a realizar la operacion 3 - 3 + 3 para obtener una imagen cancerosa con estructuras tubulares.
In [75]:
normal_tub_1 = np.loadtxt(FULL_OUTPUT_PATH + 'histo__57_19.txt')
normal_tub_2 = np.loadtxt(FULL_OUTPUT_PATH + 'histo__75_24.txt')
normal_tub_3 = np.loadtxt(FULL_OUTPUT_PATH + 'histo__21_7.txt')
normal_no_tub_1 = np.loadtxt(FULL_OUTPUT_PATH + 'histo__40_30.txt')
normal_no_tub_2 = np.loadtxt(FULL_OUTPUT_PATH + 'histo__9_3.txt')
normal_no_tub_3 = np.loadtxt(FULL_OUTPUT_PATH + 'histo__20_2.txt')
cancer_no_tub_1 = np.loadtxt(FULL_OUTPUT_PATH + 'histo__15_11.txt')
cancer_no_tub_2 = np.loadtxt(FULL_OUTPUT_PATH + 'histo__18_6.txt')
cancer_no_tub_3 = np.loadtxt(FULL_OUTPUT_PATH + 'histo__57_19.txt')
resultado_1 = normal_no_tub_1 - normal_tub_1 + cancer_no_tub_1
resultado_2 = normal_no_tub_2 - normal_tub_2 + cancer_no_tub_2
resultado_3 = normal_no_tub_3 - normal_tub_3 + cancer_no_tub_3
resultados = [normal_tub_1,normal_tub_2,normal_tub_3,
normal_no_tub_1,normal_no_tub_2,normal_no_tub_3,
cancer_no_tub_1,cancer_no_tub_1,cancer_no_tub_1,
resultado_1,resultado_2,resultado_3]
plt.figure(figsize=(10, 10))
plt.subplots_adjust(hspace=0.2, wspace=0.03)
for i,img in enumerate(resultados):
img = np.expand_dims(img,0)
z_ = tf.placeholder(tf.float32, [1, Z_NOISE_DIM])
samples = session.run(dcgan.sampler(z_,1), feed_dict={z_: img})
img = np.squeeze(samples)
subplt = plt.subplot(4, 3, i+1)
subplt.set_title("" if i<3 else "-" if (i>2 and i<6) else "+" if (i>6 and i<9) else "=").set_fontsize(20)
plt.imshow(img / 2 + 0.5)
plt.xticks([])
plt.yticks([])
In [77]:
src_dir_test= FULL_OUTPUT_PATH
model="evaluation/trained_models/incp_v3_15k_l0_real.pb"
labels="evaluation/labels.txt"
list_to_show = ['histo__57_19.png','histo__75_24.png','histo__21_7.png','histo__40_30.png','histo__9_3.png','histo__20_2.png','histo__15_11.png','histo__18_6.png','histo__57_19.png']
label_and_show(src_dir_test, model, labels, image_list = list_to_show , show = 16)
Otro tipo de operacion que podemos realizar es elegir dos tensores y avanzar por la "linea" que los une en el espacio latente, creando asi una interpolacion suave de imagenes intermedias
In [5]:
from utils import natural_keys
#clean up
[remove(file) for file in glob(FULL_OUTPUT_PATH+"*interp*.png")]
punto_a = np.loadtxt(FULL_OUTPUT_PATH + 'histo__9_3.txt')
punto_b = np.loadtxt(FULL_OUTPUT_PATH + 'histo__15_11.txt')
latent_space.interpolate(session, punto_a, punto_b, 132)
all_files = glob(FULL_OUTPUT_PATH+"*interp*")
all_files.sort(key=natural_keys)
plt.figure(figsize=(20, 18))
plt.subplots_adjust(hspace=0.2, wspace=0.03)
for i,img in enumerate(all_files):
img = plt.imread(img)
plt.subplot(11, 12, i+1)
plt.imshow(img, aspect="auto")
plt.xticks([])
plt.yticks([])
In [6]:
import imageio
images = []
for filename in all_files:
images.append(imageio.imread(filename))
imageio.mimsave('histo.gif', images)
In [ ]:
close_session(sess)